Hari YLN

A config-driven framework for LoRA adapter fine-tuning of open-source causal LLMs — enabling rapid experimentation and production deployment for classification, information extraction, and free-form generation. Training layer: dual-backend engine (Unsloth — ~2× faster, 60–80% less VRAM via QLoRA — plus HuggingFace PEFT/Transformers), single- and multi-GPU via Accelerate with automatic backend selection. Inference layer: standard PEFT adapters hot-loadable into a vLLM OpenAI-compatible server alongside a frozen base model — no separate deployment per fine-tuned model; containerised on GKE with HPA autoscaling and Istio routing. MLOps: end-to-end artefact lifecycle (model resolution local HF cache → GCS → HF Hub), automated post-training evaluation (accuracy, confusion matrix, MSE) with structured JSON results and auto-generated vLLM serve commands, and GCS artefact upload; Kubernetes manifests with PVC model caching and GMP PodMonitoring. Zero-Python task addition via YAML + prompt file + CSV; multi-task single-adapter training.

A high-scale fraud detection ecosystem that intercepts threats across the fulfilment journey — payment, refund, reship — by analysing real-time behavioural signals, clickstream, and historical patterns. Streaming: Apache Beam pipelines on GCP Dataflow processing 40k events/sec at peak, transforming raw interactions into dynamic features for the Vertex AI Feature Store. Inference: real-time endpoints on GKE holding a 200 ms scoring SLA. Training: continued training via Vertex AI Pipelines. Transitioned validated XGBoost models to production with Helm + Vertex AI Model Registry; provisioned GCP resources (Pub/Sub, Buckets, Dataflow) via Google KCC; GitOps with Cloud Build + ArgoCD; tracked experiments and model versions in a bespoke in-house MLOps portal; JWT-secured endpoints with rigorous load testing via Cloud Functions; Prometheus/Grafana + Cloud Monitoring alerting to Slack/email.

Helped Cyber Security act faster on critical threats reported to the "abuse" mailbox. Two-part solution in active use: (1) an email-routing structure that classifies inbound abuse mail with ML, strips junk, and forwards to the right action teams; (2) an extraction pipeline that pulls scammers' email addresses and phone numbers — even when hidden in HTML, images, PDFs/PNGs, multi-coloured digits, or whitespace tricks — using OCR (pytesseract) and rule-based parsing, then notifies stakeholders (mail partners such as Google and iCloud, the cyber threat-response team, and the phishing team) so partners can block the numbers and shut down the scammers.

Profiled AWS account activity from security logs to surface anomalous/nefarious behaviour. Every account, user, service, and machine gets a digital fingerprint — an unsupervised autoencoder trained per entity to learn its normal moment-by-moment activity; deviations trigger alerts, and models are continuously retrained as behaviour evolves. Engineered training and inference pipelines on NVIDIA Morpheus, exploiting GPU-accelerated libraries on an NVIDIA DGX server (8×A100, 1TB RAM). Operating these per-account pipelines means handling enormous data — ~20 days across 3 AWS accounts can exceed a billion rows — with high feature cardinality. Predictions served to the cyber team via Grafana; models managed with MLflow.

Training and inference pipelines that score phishing probability and analyse trends in incoming mail — augmenting existing Proofpoint defences by catching attempts that slipped past initial filters and accelerating quarantine. Data moved from Elastic Search to Microsoft Sentinel; Morpheus pipelines run multi-stage on a GPU server (raw download in a time-delta loop → pre-processing → featurization → inference → alerting), with the training pipeline folding in security-team feedback for continuous improvement. Detections let the security team pull mail from inboxes, warn employees, or open investigations.

Optimised the last mile — transporting pallets from a depot to many grocery-chain stores by distribution truck. Replaced manual, ad-hoc route planning with an OR-Tools Vehicle Routing Problem formulation honouring real constraints: number of vehicles, store time windows, waiting times, and demand fulfilment for every store. Computed key performance metrics, compared generated vs. manually-planned routes, and ran many input scenarios to analyse model behaviour. Used Google Maps services for coordinates/directions; built and analysed in Python + Jupyter with GitHub version control.

Designed and built a simulation framework modelling the movement of products from vendors and producers to customers and intermediate warehouses, with an embedded optimisation model for the flow of goods across every touchpoint. Tracked KPIs such as operational cost, product freshness, and carbon footprint; ran scenario simulations that informed critical business decisions. Built robust pipelines to feed simulation results into Power BI and managed MLOps for smooth deployment and monitoring.

Minimised transportation costs for daily transport between a city's terminals and distribution centres for Norway's largest fruit & vegetable distributor. Replaced a highly-manual Excel process ("ProductTransport") with an OR-Tools MIP assignment model. The transport model was hosted on Azure Function Apps behind API endpoints, fronted by a Logic App that routes requests by location; a scheduled RPA process retrieves optimised plans and Slack notifications keep stakeholders updated. CI/CD via Azure DevOps + GitHub.

A proof-of-concept analysing the impact of introducing distribution centres at strategic locations in a logistics network for Norway's largest vegetable distributor — assessing whether they could reduce volume through main warehouses, cut cost, and deliver fresher product. Collected, cleaned, and pre-processed raw network data; analysed current goods flow; designed approaches to simulate real-world scenarios; surfaced new opportunities and the customers/vendors who would benefit most; and communicated actionable recommendations to stakeholders.

Designed, implemented, and maintained an ML model to predict the quality of goods on arrival at the warehouse for Norway's largest fruit & vegetable distributor, so the inspection team could prioritise efficiently. Built end-to-end on Azure — ML Studio, Logic Apps for workflow, Pipelines for automated retraining, and Azure Storage. Owned retraining strategy (MLOps), model optimisation, client demos, performance monitoring, and ongoing support; worked closely with stakeholders to measure and track impact.

Built and deployed a forecasting system predicting company-wide data across many intelligence centres for a UK-based global business-intelligence provider. Domains spanned Consumer, Oil & Gas, Power, Financial Services, Retail, Technology, Construction, Mining, Aerospace, Defense & Security, Telecommunication, Automotive, Travel & Tourism, and more — conducting in-depth data research, generating forecasts, and adjusting them for domain-specific scenarios.

A robust hybrid on-prem/cloud computer-vision system for multi-class classification of faulty chips on the production line. Built concurrent training and inference pipelines: raw data downloaded to a GCP compute cluster via multi-processing/threading, converted to TF records, ingested and pre-processed on CPU, with training distributed across multiple GPUs and a parallel evaluation job gathering checkpoints and signalling when to stop. A DenseNet trained on ~1 million images in about 3 hours; inference served via Docker + TensorFlow Serving. Led the team on inference, database, and data-handling, deployed at the client's South Korea site, and trained the client team.

Developed an opportunity win/lose prediction system for the sales & marketing team of a US-based multinational internet-technologies company, enabling data-driven decisions through predicted probability scores. Used classical ML (logistic regression), built a data extraction/load system with SQL Server Integration Services (SSIS), and integrated the predictions into a Power BI dashboard backed by an SSAS cube and SQL Server.

Generated ID cards for new members enrolled under a leading US health-insurance client's policies — spanning estimation, impact analysis, design, coding, review, unit and integration testing across both agile and waterfall. Built on the mainframe stack (JCL, COBOL, DB2, VSAM, CICS), with post-implementation support, enhancements, and knowledge-transfer sessions for the team.

Interactive tool to find the optimal vLLM-compatible LLM for any GPU — 115 models filtered by VRAM, quantization, and KV-cache fit across A100 → B200.

Daily AI news brief from free sources (RSS, arXiv, Hacker News), synthesised with Gemini and exported as a polished PDF + Markdown.

Run any LLM fully locally on a Mac with a single command — private, offline, and zero cloud dependency.

Remote command executor & project launcher — single binary, zero dependencies, real-time SSE streaming.

A self-contained personal & family finance tracker that runs entirely on your machine — no accounts, no cloud, no dependencies. Just Python and a browser.

Building LLM inference internals from scratch — KV-cache, paged attention, and speculative decoding — alongside a config-driven LoRA → eval → vLLM-serve pipeline. Shipping in public through 2026.